Table of Contents
- Key takeaways
- What is chatbot design?
- Chatbot design principles from usability research
- The 8 chatbot design patterns
- 1. Scope: decide what the chatbot handles
- 2. Persona and tone: decide how the chatbot sounds
- 3. Opening turn: design the first message
- 4. Input: offer buttons and free text together
- 5. Answers: put the answer first and keep it short
- 6. Flow and context: let visitors change course
- 7. Fallback and handoff: design the misses
- 8. Widget UI: design the chat window
- How to design friendly chatbot responses
- The chatbot design process in six steps
- Scripted flow, AI answers, or both?
- Worked example: a design spec for a store chatbot
- How to test a chatbot design
- Common chatbot design mistakes
- Frequently asked questions
- Basics
- What is chatbot design?
- What are chatbot design patterns?
- What is the difference between chatbot design and conversation design?
- Opening messages and tone
- What should a chatbot's first message say?
- How do you write a friendly chatbot greeting that is still concise?
- Should a chatbot have a personality?
- How should a chatbot handle small talk?
- Flows and fallback
- Should a chatbot use buttons or free text?
- How many times should a chatbot retry before handing off?
- Do AI chatbots still need conversation design?
- Interface and testing
- Where should a chat widget go on a page?
- What accessibility rules apply to a chat widget?
- How do you test a chatbot design before launch?
- Sources
Created by
Do not index
Created time
Sep 27, 2023 11:12 PM
Chatbot design is the work of deciding what a chatbot does, how it talks, and how its chat window behaves, and it comes down to eight patterns: scope, persona, opening turn, input, answers, flow, fallback, and widget UI.
TL;DR
- Design the scope first. Nielsen Norman Group's guidelines say to tell people which tasks the bot can do and to avoid false expectations.
- Write the opening turn with care. It should say the assistant is a bot, name three or four topics, and show them as buttons.
- Friendly means clear and consistent. Pick 4 to 6 persona adjectives and hold the same tone from the greeting to the last reply.
- Offer buttons and free text together. Rephrase once after a miss, then offer a person.
- Check the widget against WCAG 2.2: 4.5:1 text contrast, full keyboard operation, and tap targets of at least 24 by 24 CSS pixels.
Key takeaways
Pattern | Design decision | What good looks like |
1. Scope | Which tasks the bot handles | A written list of in-scope and out-of-scope topics |
2. Persona and tone | How the bot sounds | 4 to 6 adjectives and one consistent tone |
3. Opening turn | What the first message says | Bot disclosure, three or four topics, starter buttons |
4. Input | How visitors reply | Buttons for common choices, free text always open |
5. Answers | How replies are shaped | Answer in the first sentence, details on request |
6. Flow and context | How the conversation moves | Visitors can change course, and nothing is asked twice |
7. Fallback and handoff | What happens after a miss | One rephrase, then a route to a person |
8. Widget UI | How the chat window behaves | One entry point on every page that passes WCAG 2.2 AA |
What is chatbot design?
Chatbot design is the set of decisions that shape a conversation between a visitor and a bot before any building starts. It has two halves. Conversation design covers what the bot says and when. Interface design covers the chat window the visitor sees and clicks.
The two halves need each other. A well-written greeting fails inside a widget nobody can find. A polished widget fails when the first message promises help the bot cannot give.
This guide covers design. Running a chatbot after launch is a different job, and the chatbot best practices guide covers it with 15 operating rules. Picking a canvas tool to draw flows is covered in the guide to drag-and-drop chatbot builders.
Chatbot design principles from usability research
Chatbot design principles come from watching real people use bots. Nielsen Norman Group ran a usability study of chatbots in 2018 with 8 US participants on mobile and desktop. The study ends with nine guidelines, and five of them decide most designs:
- Be upfront that the visitor is talking to a bot.
- Tell people clearly which tasks the bot can do.
- Let people use both free text and selectable links.
- Save information from one task to the next.
- Be honest about not understanding, and offer a way out to a person.
The same group published 10 guidelines for site AI chatbots in April 2026, based on users working with AI chatbots on several sites. Those guidelines focus on the chat window: one entry point, presence on every page, suggested prompts as buttons, images in answers, and no autoscroll.
Google's conversation design guide adds one warning about errors: "One poorly handled error can outweigh dozens of successful interactions." Microsoft's Bot Framework guidance sets the test for the whole design. A good bot solves the visitor's problem with the fewest back-and-forth turns.
The 8 chatbot design patterns
The eight chatbot design patterns below follow the order of a real design process. Each one is a decision to write down before building.
1. Scope: decide what the chatbot handles
The scope pattern is a written list of tasks the bot handles and tasks it sends elsewhere. Scope comes first because every later pattern depends on it.
Nielsen Norman Group's guidance is to build bots for simple tasks, because complex ones are handled poorly in a small chat interface. Their study also showed the cost of an unclear scope. Participants could not tell which drugs a health bot knew about, so they could not judge whether it would help.
- Write down: five to ten in-scope topics, and the topics that always go to a person.
- Design check: every in-scope topic has content the bot can answer from.
2. Persona and tone: decide how the chatbot sounds
The persona pattern is a short description of who the bot is, used to keep every reply in one voice. Google's conversation design guide describes a persona as a design tool for writing conversations. It notes that users project a persona onto a bot whether one was planned or not.
Google's method takes four steps:
- Brainstorm adjectives that describe how the bot should come across.
- Narrow the list to 4 to 6 key adjectives.
- Think of a few characters that fit those adjectives.
- Pick one character and describe it in a paragraph at most.
The guide also says the goal of a persona is never to make people think they are talking to a human.
Keep the tone consistent. In Nielsen Norman Group's study, a playful insurance bot dropped its tone once it started collecting quote details. One participant said the change made the personality feel misleading.
- Write down: the adjectives, three sample replies, and words the bot never uses.
- Design check: read the greeting, an answer, and the fallback message together. They should sound like one speaker.
3. Opening turn: design the first message
The opening turn pattern is the first message plus the buttons shown with it. It does four jobs: it says the assistant is a bot, names what it covers, offers starter topics, and leaves the text box open.
Microsoft's guidance is direct on this point. It says a bot should never start with an open-ended question such as "How can I help you?", because visitors cannot guess what the bot does. Nielsen Norman Group's 2026 study found the same problem. One chatbot opened with a vague greeting that promised help with anything, and participants found its limits the hard way.
Three opening patterns cover most sites.
Opening pattern | Example first message | Use it when |
Menu opening | "Hi, this is the Acme assistant, an AI. Pick a topic or type a question." Buttons: Shipping, Returns, Sizing, Talk to a person | The bot covers a few clear topics |
Task opening | "Hi, this is the Acme booking assistant, an AI. Which day would you like to come in?" | The bot does one job |
Page-aware opening | On the pricing page: "Questions about plans? Ask here." Buttons: Compare plans, Annual discount, Cancel anytime? | The site has pages with different intent |
Nielsen Norman Group recommends the page-aware version where possible. In its study, Amazon's assistant changed its suggested questions on product pages, which showed visitors that the bot knew what they were viewing.
- Write down: the greeting text, three or four starter buttons, and any page-specific versions.
- Design check: a new visitor can say what the bot does after reading the opening turn once.
4. Input: offer buttons and free text together
The input pattern gives visitors two ways to reply: buttons for common choices and a text box for everything else. Nielsen Norman Group found that both must be present. Buttons save typing, and free text lets people leave the script.
The study has examples of each failure. One participant was annoyed at having to type a date when a button would have done. Another bot removed the text box after a menu choice, which left visitors with no way to ask anything else.
The 2026 guidelines extend this to follow-up questions. Suggested questions should be clickable buttons, shown at the start and again after each answer. Repeating a suggestion the visitor already skipped makes the bot feel pushy. When a step collects fixed details such as an address or a card number, a form may suit it better than chat, and the chatbots vs forms guide has a decision matrix for that choice.
- Write down: which steps get buttons, and the wording of each one.
- Design check: every button step also accepts typed input.
5. Answers: put the answer first and keep it short
The answer pattern sets the shape of every reply: the answer in the first sentence, a few lines at most, and more detail only when asked. Long replies push earlier content out of view in a small window.
Nielsen Norman Group's 2026 guidelines add three rules for answers:
- Use progressive disclosure. Let visitors expand and collapse extra detail in place.
- Include images. Product suggestions and how-to steps are easier to judge with a picture.
- Do not autoscroll to the end of a long reply. Visitors then have to scroll back up to start reading.
For AI chatbots, the answer pattern also includes a source link, so visitors can check where an answer came from. The guide to stopping chatbot hallucinations covers the guardrails behind that.
- Write down: a length limit, a format for steps, and whether sources are shown.
- Design check: the first sentence of a sample reply answers the question on its own.
6. Flow and context: let visitors change course
The flow pattern maps the steps of each task and decides what the bot remembers between them. Nielsen Norman Group describes most bot tasks as linear flows with a few branches. The flows worked while visitors gave expected answers and broke when they did not.
Two findings shape the design. One bot forced a participant to restart a whole decision tree after a single unexpected answer. Another asked for an address twice in one session. The guidelines that follow are to allow jumps forward and backward, and to carry information from one task to the next.
Microsoft's navigation guidance names five bot failures to design against.
Failure | What the bot does | Design fix |
The stubborn bot | Keeps asking the same question when the visitor changes topic | Let visitors change course, and cap retries per question |
The clueless bot | Gives a nonsense reply to "help" or "cancel" | Handle help, cancel, and start over at every step |
The mysterious bot | Shows nothing while it works | Acknowledge the message at once |
The captain obvious bot | Offers obvious information nobody asked for | Give only information the visitor can use |
The bot that can't forget | Brings up past conversations that are not relevant | Stay on the current topic until the visitor returns to an old one |
Small talk belongs in the flow design too. A workable pattern is to answer a greeting or a thank-you in one line and return to the task. A bot that chats at length delays the answer the visitor came for.
- Write down: one sample dialog per task, including one where the visitor changes their mind.
- Design check: in every sample dialog, the bot never asks for a detail it already has.
7. Fallback and handoff: design the misses
The fallback pattern decides what the bot says when it does not understand or has no answer. Every bot misses sometimes, so the miss needs a design of its own.
Google's error guidance gives a clear sequence:
- On the first miss, ask again in different words. Never repeat the same prompt verbatim.
- On the next miss, add support such as options or an example.
- After two failed attempts to collect the answer, stop asking and tell the visitor another way to finish the task.
On a website, that other way is usually a person, and the offer can appear from the second miss. Nielsen Norman Group's study found that participants preferred an honest "I don't understand" to a wrong answer. They reacted well when a bot admitted the failure and offered a phone number or a live agent.
Situation | Fallback message design |
First miss | Rephrase the question and show two or three options as buttons |
Second miss | Add an example or options, and show a route to a person |
Third miss | Stop asking, say the bot cannot help with this, and hand off |
Topic is out of scope | Say so in one line, name what the bot does cover, and offer a person |
Visitor asks for a person | Hand off at once and say what happens next |
This guide covers the wording and placement of the handoff. The rules for when a bot should hand off are in the chatbot best practices guide. The comparison of AI chatbots with human handoff shows how much context each tool passes to the agent.
- Write down: the fallback messages, the retry limit, and the handoff confirmation text.
- Design check: a person is never more than two failed attempts away.
8. Widget UI: design the chat window
The widget UI pattern covers where the chat button sits, how the window behaves, and whether everyone can use it. Nielsen Norman Group's 2026 guidelines and WCAG 2.2 supply most of the rules.
Element | Design rule | Source |
Entry point | One chat entry point per site. Two competing chat widgets confuse visitors. | Nielsen Norman Group, 2026 |
Availability | Keep the chat on every page of a task, including checkout | Nielsen Norman Group, 2026 |
Window size | Let visitors resize the window for long or visual answers | Nielsen Norman Group, 2026 |
Text contrast | At least 4.5:1 for normal text | WCAG 2.2, criterion 1.4.3 |
Keyboard | Every function works from a keyboard | WCAG 2.2, criterion 2.1.1 |
New messages | Assistive technology can announce them without moving focus | WCAG 2.2, criterion 4.1.3 |
Tap targets | At least 24 by 24 CSS pixels | WCAG 2.2, criterion 2.5.8 |
Small screens | Content reflows at 320 CSS pixels wide with no sideways scrolling | WCAG 2.2, criterion 1.4.10 |
Brand colors are the usual cause of a contrast failure. White text on a light brand color often falls below 4.5:1, so check the pair before launch.
- Write down: button position, colors with their contrast ratio, icon, and window behavior on mobile.
- Design check: complete one full conversation with only a keyboard.
How to design friendly chatbot responses
Friendly chatbot responses are clear, direct, and consistent. A reply that saves the visitor time reads as friendly. Extra warm wording adds length and delays the answer.
Four rules cover most replies:
- Acknowledge the question in a few words, then answer it.
- Use plain words and the visitor's own terms.
- Offer one next step.
- Keep the same tone when the news is bad.
Situation | Unfriendly | Chatty | Friendly and concise |
Greeting | "State your query." | "Hey there, superstar! So happy you stopped by today!" | "Hi, this is the Acme assistant, an AI. Ask about orders, returns, or sizing." |
Answer | "See the returns policy." | "Great question! Returns are something lots of people ask about, so here goes..." | "You can return any item within 30 days. Want the steps?" |
Fallback | "Invalid input." | "Oops! Silly me, my circuits are tangled again!" | "That one is outside what this assistant covers. Would you like to talk to a person?" |
Handoff | "Transferring." | "Hang tight while the humans get rounded up!" | "Passing this to the team with your messages so far. They reply by email." |
Nielsen Norman Group observed that people who know they are talking to a bot write more directly and drop politeness markers. A friendly bot can match that. Short, direct replies read as helpful in chat.
Greetings follow the same rule. A friendly greeting fits in two short sentences: one that says who is speaking, and one that says what the visitor can ask.
The chatbot design process in six steps
The chatbot design process runs from scope to testing in six steps.
- Collect real questions. Pull the last 100 questions from email, chat, and site search.
- Set the scope. Sort the questions into topics the bot handles and topics for a person.
- Write the persona. Choose the adjectives and write three sample replies.
- Write sample dialogs. Google's guide recommends sample dialogs as a way to try design ideas before building. Write one per task, plus one where the visitor goes off script.
- Design the opening turn, fallback messages, and widget. Use patterns 3, 7, and 8.
- Test with real people, then revise. Watch a handful of people try the main tasks and note where they stall.
Preparing the content the bot answers from is a separate step, covered in the guide to chatbot training.
Scripted flow, AI answers, or both?
The first design choice is how the bot produces its replies, because it decides how much of the conversation the team writes by hand. A scripted flow needs every step designed. An AI chatbot trained on existing content needs the scope, persona, opening turn, and fallback designed, and it writes the answers itself.
If you need | Pick | Why |
Answers to open questions from a website, help center, or documents | A knowledge-trained AI chatbot | Nobody can script every phrasing, and the answers already exist in the content |
A fixed sequence such as a booking, a quote, or lead qualification | A scripted flow in a visual builder | Every step is predictable and can be tested |
Both open questions and one or two fixed sequences | A hybrid of the two | Each part of the conversation gets the method that suits it |
Chat and voice agents designed step by step by a dedicated team | A conversation design platform | It gives designers control over every turn |
Refunds, cancellations, or anything that commits money | A person or a fixed flow | A generated reply should not make the commitment |
Visual flow tools are compared in the guides to drag-and-drop chatbot builders and no-code chatbot builders. For the step-by-step platform route, see the Voiceflow review.
SiteGPT is a knowledge-trained option. It has no visual flow canvas, so it is the wrong pick for a bot that is mostly scripted steps. Its settings map to the patterns in this guide:
Pattern | SiteGPT setting |
Persona and tone | 12 built-in personas, plus custom instructions |
Opening turn | A welcome message, and conversation starter buttons that can be limited to chosen pages |
Input | AI-written follow-up suggestions after each answer, 3 by default |
Fallback and handoff | Human support with escalation buttons, on any plan |
Widget UI | Colors, icon, position, dark mode, and font size |
Starter costs $59 a month billed monthly, or $39 a month billed yearly. There is a 7-day free trial that requires a credit card.
Worked example: a design spec for a store chatbot
This example uses a made-up online bike shop to show what a finished chatbot design looks like on one page. The details are illustrative.
Pattern | Decision |
Scope | In: order status, shipping, returns, sizing. Out: refunds, warranty claims, repairs |
Persona | Knowledgeable, plain-spoken, patient, brief |
Opening turn | "Hi, this is the Pedal assistant, an AI. Ask about orders, shipping, returns, or sizing." Four starter buttons |
Input | Buttons for the four topics and for frame sizes. Free text open at every step |
Answers | Three sentences at most, a source link, a size chart image for sizing questions |
Flow | Order number asked once and reused. "Start over" works at any step |
Fallback | One rephrase with buttons, then "Talk to a person" |
Widget | Bottom right on every page. Dark blue button with white text |
The team then scores the design against two checks per pattern, 16 checks in total.
Result | Checks | Share |
Passed | 13 | 81% |
Failed | 3 | 19% |
Total | 16 | 100% |
The three failures are specific. The opening turn does not change on product pages. The order number is asked again after a handoff. The starter buttons are 20 pixels tall, under the 24-pixel minimum. Fixing those three brings the score to 16 out of 16 before any visitor sees the bot. To turn a spec like this into instructions a bot can follow, start from a chatbot template and fill in the decisions from the table.
How to test a chatbot design
Testing a chatbot design means watching people use it and checking each pattern against a pass mark. Nielsen Norman Group drew the nine guidelines in its 2018 study from sessions with 8 participants, so a test does not need a large group.
Test | How to run it | Pass mark |
Scope test | Show five people the opening turn for ten seconds | Each can name two things the bot does |
Off-script test | Give an unexpected answer in the middle of each flow | The bot recovers without a restart |
Memory test | Start a second task in the same chat | The bot does not ask again for known details |
Fallback test | Ask three out-of-scope questions | The bot declines and offers a person each time |
Tone test | Read the greeting, an answer, and the fallback aloud | They sound like one speaker |
Keyboard test | Complete a conversation without a mouse | Every button and field is reachable |
Contrast test | Check the text and button colors | 4.5:1 or higher for normal text |
Design testing happens before launch. The fixed question set and the weekly numbers to track after launch are in the chatbot best practices guide.
Common chatbot design mistakes
The most common chatbot design mistakes each break one of the eight patterns.
- An open-ended greeting. Visitors ask for things the bot cannot do, and the first reply is a failure.
- A personality that fades. The greeting is playful, and every later message is flat.
- Buttons with no text box. Visitors with a question outside the menu have nowhere to go.
- A fallback that loops. The bot repeats the same question until the visitor leaves.
- Two chat widgets on one site. Visitors cannot tell which one to use.
- Brand colors that fail contrast. Some visitors cannot read the replies.
Frequently asked questions
Basics
What is chatbot design?
Chatbot design is the work of deciding what a chatbot handles, how it talks, and how its chat window looks and behaves. It covers scope, persona, the opening message, input options, answer format, conversation flow, fallback messages, and the widget interface.
What are chatbot design patterns?
Chatbot design patterns are reusable solutions to problems that come up in most chatbots. Examples are an opening turn that states the bot's scope, starter buttons next to a text box, and a fallback that rephrases once and then offers a person.
What is the difference between chatbot design and conversation design?
Conversation design is one half of chatbot design. It covers the words and the order of the conversation. Chatbot design also includes the interface: the chat button, the window, colors, and accessibility.
Opening messages and tone
What should a chatbot's first message say?
A chatbot's first message should say that the assistant is automated, name three or four topics it covers, and show those topics as buttons. Microsoft's guidance warns against an open-ended first question, because visitors cannot guess what the bot does.
How do you write a friendly chatbot greeting that is still concise?
Write two short sentences. The first says who is speaking and that it is an AI. The second says what the visitor can ask about. Put the topics on buttons so the greeting itself stays short.
Should a chatbot have a personality?
A chatbot should have a consistent persona. Google's conversation design guide recommends 4 to 6 adjectives and a one-paragraph description. The persona should never suggest the bot is human, and the tone should hold for the whole conversation.
How should a chatbot handle small talk?
A chatbot should answer small talk in one line and return to the task. A greeting gets a greeting and a reminder of what the bot covers. A thank-you gets a short reply and an offer of further help.
Flows and fallback
Should a chatbot use buttons or free text?
A chatbot should use both. Nielsen Norman Group found that buttons save typing for common choices, and free text lets visitors ask questions outside the script. Removing either one caused problems in its study.
How many times should a chatbot retry before handing off?
Two failed attempts is a sound limit. Google's error guidance says to rephrase on the first miss, add support on the next, and end after two failed attempts to collect the answer. On a website, offer a person from the second miss.
Do AI chatbots still need conversation design?
AI chatbots need less scripting and the same design decisions. The AI writes the answers, and the team still designs the scope, persona, opening turn, fallback messages, handoff, and widget.
Interface and testing
Where should a chat widget go on a page?
A chat widget should sit in one consistent place on every page of a task. Nielsen Norman Group's 2026 guidelines say to keep the chatbot available across pages and to merge separate chat features into one entry point.
What accessibility rules apply to a chat widget?
WCAG 2.2 is the standard to test against. The key criteria are text contrast of at least 4.5:1, full keyboard operation, status messages that assistive technology can announce, and tap targets of at least 24 by 24 CSS pixels.
How do you test a chatbot design before launch?
Test a chatbot design by watching a handful of people try the main tasks. Check that they understand the scope from the opening turn, that the bot recovers from unexpected answers, and that a person is reachable after two misses.
Sources
Every source below was opened on 8 October 2026.
Written by
